LLM’s are not the end all be all. What other AI tech are you all using? Something generative? Something else?
It’s kinda fun to poke around with image and video generation using ComfyUI. There’s all sorts of models, LORAs, and setting to tune and play with
I want to but I don’t think I have the hardware to support it. Need at least a decent GPU, right?
For playing around, no. My GPU is 9 years old and has 8GB VRAM. I generating my 384x512 profile picture with KoboldCPP (Vulkan) and Z-Image-Turbo model locally.
I just regenerated my profile picture with KoboldCPP a bit differently. I used some clearer LoRA configurations than I did initially. For Z-Image-Turbo it’s CFG 4 and for Flux2 Klein 4b CFG 1, both used Euler. For Anime V1 LoRA I started with
anime illustrious,and for Koni Animestyle LoRA withanime_style,. :384×512 & Z-Image-Turbo & Anime V1 & 8 steps (2min 42s):
384×512 & Z-Image-Turbo × Anime V1 & 25 steps (8min 16s):
384x512 × Flux2 Klein 4b × Koni Animestyle × 4 steps (22s):
384x512 × Flux2 Klein 4b × Koni Animestyle × 8 steps (35s):
384x512 × Flux2 Klein 4b × Koni Animestyle × 25 steps (1min 28s):
Upscaling it with RealESRGAN x4plus anime 6B upscaler only takes 10 seconds longer with this dimension.
KoboldCPP outputs how long generation takes like:
[11:09:15] Generating Image (8 steps) |==================================================| 480/480 - 0.00MB/s |==================================================| 480/480 - 0.00MB/s |==================================================| 480/480 - 810.94MB/s |==================================================| 480/480 - 0.00MB/s |==================================================| 8/8 - 19.64s/it [11:11:57] Generating Media CompleteThose are pretty good results! Unfortunately, I don’t have a GPU at all yet. Any recommendations on something capable yet affordable? What would you buy next?
I stick to what I have as long as it continues to work for me, therefore I haven’t looked into GPUs the last years. New Radeon RX 580 is still at 200€, I paid 300€ 8 years ago. Looks like RX 580 2048SP 8GB is actually one of the cheapest one can get currently, but not entirely sure. (My GPU has 2304 shading units.) It’s so old that ROCm support was dropped and you really need software that can do AI via Vulkan
edit: looks like Intel Arc A750 is 5 years younger, has better performance and more memory bandwidth
Thanks! I’ll make sure to include Vulcan in criteria for a new machine and I’ll look at intel options. Only been looking at NVIDIA. Even 4GB VRAM will be a huge upgrade for me.
All modern GPU drivers support Vulkan, at least on Linux. I don’t know how bad Vulkan does in comparison to CUDA, since I’ve never used that stuff. The bigger issue is software-side support. Ollama, llama-cpp and KoboldCPP all support Vulkan by now. ComfyUI doesn’t seem to support it.
Thanks! I’ve been using llama.cpp but not married to it. What’s your opinion on used machines? Presumably risky to get a used GPU or they don’t really just fail?
Ah yeah, I think you need a 6GB gpu for images and probably like a 12Gb GPU for videos.
I think you can run with models without, but it just slows way down
Agents. And finding sane ways to run them safely.
Yeah big move towards agents. They’re based on LLM’s, no?
Yeah. Basically there like giving an LLM a set of tools it can use on its own. It’s a lot like having an intern at a company. You don’t give them access to anything you don’t want destroyed. But you can give them simple/safe buttons to push.
For sure. I was wondering what people are doing beyond LLM’s. Is there some next gen take that’s coming out now?
